Cambridge Centre for Alternative Finance

Success at the first Global Dialogue on AI Governance should be measured not by the breadth of consensus achieved, but by the depth of commitments made and whether those commitments reach the populations currently excluded from both AI development and AI policymaking. Three outcomes would mark genuine success: First, a shift from principles to standards. The global AI governance space already has no shortage of ethical principles and high-level declarations. What it lacks are binding or quasi-binding measurement standards agreed methodologies for evaluating AI performance, safety, and fairness across languages, geographies, and use cases. A successful Dialogue would produce a concrete mandate to develop such standards under multilateral auspices, with timelines. Second, meaningful inclusion of the Global South in agenda-setting. Inclusive governance cannot mean simply consulting developing nations on frameworks designed elsewhere. A successful Dialogue would establish structural mechanisms (not one-off consultations) for research communities, civil society, and governments from Africa, the MENA region, South and Southeast Asia, and Latin America to co-design the benchmarks, definitions, and red lines that will govern AI globally. Third, recognition that AI inequality is not only about access. Connectivity and hardware gaps are real, but the Dialogue should also surface a subtler and less-discussed inequity: the systematic underperformance of commercial AI in non-English languages at equivalent or higher cost. A governance framework that treats "inclusive AI" as synonymous with "connected AI" will miss the people already online, already paying, and already being underserved. A Dialogue that produces these outcomes are measurable standards, structural inclusion, and a richer definition of AI equity will have laid a foundation worth building on in 2027 and beyond.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights

These four themes, taken together, map directly onto the governance gap I focus on in my research: the systematic inequity embedded in how commercial AI systems perform across languages. Safe, secure and trustworthy AI is a priority because safety evaluation as currently practiced is overwhelmingly English-centric. A system benchmarked as safe in English may produce inaccurate, harmful, or culturally inappropriate outputs in Arabic or other languages serving hundreds of millions of users. Safety certification without multilingual validation is incomplete by definition. Social, economic, ethical, cultural, linguistic and technical implications of AI is the theme that most directly names the problem. The inclusion of "linguistic" in this framing is notable and underutilised. My research on Arabic-language AI performance and MENA adoption patterns shows that linguistic disparity is not incidental, it is a structural feature of how AI systems are built and commercialised, with real economic and social consequences for non-English-speaking populations. Transparency, accountability, and human oversight matters here because without mandatory disclosure of cross-language performance data, users and policymakers cannot see the disparity, let alone address it. Transparency must extend beyond algorithmic decision-making to include how performance varies by language, and how pricing relates to that performance. Protection and promotion of human rights grounds the other three in a normative framework. Language is not merely a technical variable, it is a dimension of identity, dignity, and access to opportunity. When AI systems systematically serve some language communities worse than others, this is not a product limitation to be tolerated; it is a human rights concern to be governed. These themes are not separate priorities for me, they are four dimensions of the same problem.

Yes. There is a structural issue that cuts across all the listed themes but is not named explicitly in any of them: the absence of cross-language performance standards for commercial AI systems. Current governance discussions address bias, safety, and accountability but almost always within a single-language or English-first frame. The result is that AI governance frameworks are being designed around systems as they perform in English, not as they perform for the majority of the world's population. This gap has several concrete dimensions that the Dialogue should treat as an emerging governance priority: Benchmarking asymmetry. The datasets and evaluation frameworks used to certify AI systems as accurate, safe, or fair are heavily weighted toward English and a small number of high-resource languages. There are no internationally agreed standards for what constitutes acceptable AI performance in Arabic, Swahili, Bengali, or the hundreds of other languages in which AI products are commercially deployed. Pricing without parity. Commercial AI APIs charge comparable rates across languages while delivering measurably inferior performance for low-resource languages. This is invisible in current governance frameworks because no disclosure requirements exist. Users in the Global South are, in effect, subsidising the development of systems optimised for wealthier, English-speaking markets. Measurement capacity in the Global South. Even where researchers wish to document these disparities, they often lack access to the proprietary benchmarking infrastructure needed to do so rigorously. Governance frameworks should actively fund and support multilingual AI evaluation capacity in underrepresented regions. I would recommend that the AI Dialogue formally recognise linguistic equity in AI performance and pricing as a cross-cutting governance issue and mandate the development of international standards for cross-language evaluation and disclosure as part of its 2026-2027 workplan.

The governance gaps identified in my selected themes have direct, measurable consequences for the MENA region and for the Arabic-speaking population more broadly, both as AI users and as a largely absent voice in AI standard-setting. The challenge of invisible underperformance. Arabic is spoken by over 400 million people, yet commercial AI systems consistently underperform in Arabic compared to English across core tasks: translation accuracy, speech recognition, natural language understanding, and generative quality. Because no international disclosure standards exist, this disparity is not surfaced in procurement decisions, regulatory assessments, or public accountability mechanisms. Governments and institutions in the region are adopting AI products without the tools to evaluate whether those products are fit for their linguistic and cultural context. Economic cost with no accountability mechanism. My research on AI adoption benchmarking across the MENA region shows that Arabic-language users pay comparable API pricing for inferior outputs. This represents a transfer of value from the Global South to AI developers in the Global North, with no governance mechanism to name, measure, or address it. It also distorts AI adoption patterns — discouraging use cases where linguistic underperformance would be visible, such as public services, legal systems, and education. A structural absence from governance design. The MENA region is almost entirely absent from the committees, consortia, and research institutions that design AI benchmarks and safety standards. This means the metrics by which AI systems are evaluated as "safe" or "fair" do not reflect Arabic linguistic norms, cultural contexts, or regional risk profiles. Governance frameworks built on these standards will export that absence globally.

The AI Dialogue occupies a unique position in the international governance landscape: it is the first UN-mandated forum explicitly designed to be inclusive of all stakeholder categories and all regions, not just those with the most advanced AI industries. That positioning is its greatest asset — and its greatest responsibility. The Dialogue can advance international cooperation in three specific ways that existing bilateral or regional mechanisms cannot easily replicate. Establishing shared measurement infrastructure. Many AI governance disagreements are, at their core, disagreements about what to measure and how. The Dialogue can play a convening role in developing internationally agreed evaluation standards — for safety, fairness, and performance — that apply across languages and geographies, not just the contexts prioritised by major AI developers. This kind of standards work requires multilateral legitimacy that only a UN-anchored process can provide. Creating accountability without enforcement. The Dialogue is not a regulatory body, and it should not pretend to be one. But it can establish disclosure norms — expectations that commercial AI providers report performance data disaggregated by language, geography, and use case. Voluntary disclosure frameworks, when anchored in a UN process, carry normative weight that industry-led initiatives do not. Bridging the governance capacity gap. Many countries in the Global South lack the technical and institutional capacity to engage meaningfully in AI governance — not because they lack relevant experience or insight, but because they lack access to the data, benchmarks, and forums where governance decisions are made. The Dialogue can fund, structure, and mandate capacity-building that is directly linked to participation in governance design, not treated as a precondition for it. In short, the Dialogue's added value lies in doing what markets and major powers will not do voluntarily: centering the governed, not just the governors.

Several existing initiatives provide foundations the Dialogue should connect with rather than duplicate. UNESCO's Recommendation on the Ethics of AI (2021) is the most broadly adopted multilateral AI ethics framework to date, with 193 member states. The Dialogue should treat it as a normative baseline and focus on operationalising its commitments — particularly those on cultural diversity and linguistic pluralism — rather than re-litigating foundational principles. The OECD AI Policy Observatory provides the most comprehensive comparative dataset on national AI strategies and policies. However, its coverage skews heavily toward OECD member states. The Dialogue should both draw on this resource and mandate its expansion to include systematic data from Africa, the MENA region, and South and Southeast Asia. The Global Partnership on AI (GPAI) has produced substantive technical work on responsible AI, but its membership and working group composition remain concentrated in the Global North. The Dialogue can add value by explicitly linking GPAI outputs to implementation pathways for underrepresented regions. The ACL Anthology and multilingual NLP research communities represent a body of peer-reviewed, technical work on language model performance across languages that is directly relevant to governance but rarely cited in policy forums. Bridging this research-policy gap — by creating mechanisms for technical researchers, including those from the Global South, to inform standard-setting — would be a distinctive contribution. What the Dialogue adds that none of these provide: a single, UN-anchored forum where measurement standards, disclosure norms, and participation structures can be agreed upon across stakeholder categories and regions simultaneously. The risk is fragmentation; the opportunity is coherence. The Dialogue should explicitly position itself as the connective tissue between existing initiatives, not a replacement for them.

The format and structure of the AI Dialogue will determine whether its inclusivity commitments are substantive or ceremonial. Based on participation patterns in comparable multilateral processes, I offer the following recommendations. Structured stakeholder tracks with real agenda influence. Multi-stakeholder formats often grant civil society, academia, and technical communities access to sessions without granting them influence over agendas or outcomes. The Dialogue should establish formal mechanisms — not just open-floor periods — through which non-governmental stakeholders can propose agenda items, table draft recommendations, and respond to government positions in writing before sessions conclude. Pre-Dialogue written consultation with genuine uptake. This form is a positive step. But submissions should be systematically analysed, synthesised by independent rapporteurs, and fed visibly into the session agendas — with Co-Chairs explicitly referencing submission themes in their opening and closing remarks. Transparency about how inputs shaped the agenda would significantly increase trust and future participation. Working groups between sessions. The Dialogue currently meets in July 2026 and 2027. The intervening period should be used productively through issue-specific working groups on themes such as multilingual benchmarking, disclosure standards, and capacity-building — with open, documented membership and public outputs. Hybrid and asynchronous participation options. Meaningful participation from the Global South is constrained not only by visa access and travel costs, but by time zones, institutional resources, and language barriers in the session itself. Simultaneous interpretation, recorded sessions with transcript access, and asynchronous written contribution windows would meaningfully expand who can engage. Avoid the "consultation trap." The most common failure mode of inclusive governance processes is consulting broadly but deciding narrowly. The Dialogue should establish from the outset which decisions will be genuinely co-produced with stakeholders and which will remain intergovernmental — and be transparent about that distinction.

The most significant underrepresentation in global AI governance is not simply geographic — it is structural. Certain communities are absent not because they have nothing to contribute, but because the formats, languages, and institutional prerequisites of participation were not designed with them in mind. Arabic, Swahili, Bengali, and other non-Latin-script language communities are almost entirely absent from AI benchmark design, safety evaluation, and governance standard-setting. Their exclusion is self-reinforcing: AI systems perform worse in their languages partly because they were not involved in defining what good performance looks like. Including them requires more than translation — it requires co-designing the evaluation frameworks themselves. Independent researchers and academics from the Global South often produce directly relevant work — on language fairness, adoption patterns, risk profiles — but lack the institutional affiliations, travel funding, or professional networks to access the forums where that work could influence policy. The Dialogue should create a dedicated pathway for individual researchers, not just institutional representatives, to contribute technical expertise. Public sector practitioners in low- and middle-income countries — those actually procuring and deploying AI in education, healthcare, and public administration — carry firsthand knowledge of where governance frameworks fail in practice. They are rarely present in governance design processes. Regional pre-consultations, conducted in local languages and feeding directly into the Geneva and New York sessions, would be a practical mechanism for including them. Civil society organisations working on digital rights in the Global South have documented AI harms that do not appear in academic literature or corporate transparency reports. Their evidence base should be formally admissible in the Dialogue's deliberative processes. Inclusion, in short, requires not just opening the door but redesigning the building — starting with who gets to say what counts as a governance problem in the first place.

The most effective engagement formats are those that generate outputs — not just conversations — and that create ongoing accountability between sessions rather than concentrating all participation in high-pressure plenary moments. Evidence clinics. Rather than panel presentations, structured sessions in which researchers and practitioners present documented governance failures or successes to a panel of government and intergovernmental delegates — with delegates required to respond on record — would shift the dynamic from consultation to dialogue. These could be piloted in Geneva 2026 with outputs published as part of the official record. Red-teaming sessions on draft frameworks. Before governance recommendations are finalised, structured adversarial review sessions — in which stakeholders from underrepresented regions are specifically tasked with identifying where draft standards would fail in their contexts — would surface blind spots that conventional consultation misses. This is standard practice in AI safety; it should be standard practice in AI governance design too. A public, living repository of governance gaps. The Dialogue could maintain a continuously updated, publicly accessible document mapping identified governance gaps to proposed responses and their current status. This would make progress visible, hold stakeholders accountable between sessions, and give new participants a clear entry point rather than requiring them to reconstruct context from scratch. Language-specific breakout tracks. Running parallel sessions in Arabic, French, Spanish, and other UN languages — not just as translation services but as substantive working spaces with their own rapporteurs and outputs — would signal that linguistic diversity is a governance principle, not just a logistical accommodation. Longitudinal participation incentives. Engagement that spans both the 2026 and 2027 sessions, with recognition for sustained contributions, would build the institutional memory and trust that one-off participation cannot.

Several existing policies and practices offer concrete models the Dialogue could adapt or endorse. The EU AI Act's risk-based tiering approach provides a replicable framework for differentiating governance obligations by the severity and reversibility of potential harms. While its implementation remains primarily English-language-centric, the structural logic - proportionate obligations based on risk - is sound and could be extended to require heightened scrutiny for high-risk AI deployments in low-resource language contexts where error rates are demonstrably higher. Brazil's AI Bill and the African Union's AI Continental Strategy both represent efforts to develop governance frameworks that reflect regional values, risk profiles, and development priorities rather than simply adopting frameworks designed elsewhere. These should be treated as positive models of governance localisation, not as deviations from a global standard. The ACL Anthology's open-access publishing model - in the context of my own field - demonstrates that research infrastructure designed for global participation can produce higher-quality, more diverse outputs than closed or fee-gated alternatives. Applied to AI governance, this argues for open-access benchmarking datasets, publicly available evaluation tools, and transparent standard-setting processes. Mozilla's Common Voice project has demonstrated that community-contributed, multilingual voice data can be collected at scale through structured public participation. A similar model could underpin a UN-mandated multilingual AI evaluation corpus - publicly available, continuously expanded, and governed by a representative multilateral body rather than any single company or research institution. My own published research on Arabic speech translation and MENA AI adoption benchmarking contributes to an emerging evidence base on cross-language AI performance gaps. I would welcome the opportunity to contribute this and related work to any technical working group the Dialogue establishes on multilingual AI evaluation standards.